Investigation on Mixup Strategies for End-to-End Voice Spoof Detection System

Cross-corpora spoken language identification with domain diversification and generalization

“…For MFCC extraction, we use Hamming window of 25 ms and hop-length of 10 ms. While the mel filter banks (MFB) feature is frequently used with the neural network based classifiers, several works [37,41,81] have also applied MFCCs as input to the TDNN-based classifiers as used in our work. Note that the existing work with the three used corpora [92,53,52,11] also reported their performances mostly using MFCCs.…”

“…We have utilized the ECAPA-TDNN architecture [16], which outperformed other basic TDNN architectures for speaker verification tasks in [16]. This is also very recently adopted for language recognition tasks [37,85]. The ECAPA-TDNN model modifies the basic x-vector architecture in different ways.…”